Clef-Flash Artemis

Text-only GGUF files of Clef-Flash, a typed decision model. It returns probabilities for allowed answers through /v1/systemone; plain chat is not supported.

TurinTech models, built with Artemis

Ollama

echo "My Stripe integration keeps failing. Please help ASAP." | ollama run turintech/clef-flash-artemis:iq4_xs --ask "Does this message express urgency?"

latest selects iq4_xs. The explicit tags below use the same model bytes as the corresponding Hugging Face files. Existing size tags remain available.

GB is download size, not runtime memory. Quant labels are size classes for these mixed files; the model card lists tensor types.

For code, send model, state and typed questions to /v1/systemone. Keep questions about one input in the same request; splitting them can change the probabilities.

The package sets num_ctx 4096. Clef-Flash reads the whole input in one batch, and Ollama sizes the batch to the context for this model, so num_ctx is the longest input it accepts: 4,096 tokens, the batch size the card was measured with. That budget covers the whole encoded request (the text, the questions and the prompt format), and longer requests return an error. To accept longer inputs, build your own package with FROM turintech/clef-flash-artemis and PARAMETER num_ctx 16384 in a Modelfile. Memory use grows with the batch: in an earlier test on a 30 GB machine, a 4,914-token input stopped the Ollama runner of Clef-Flash. An options field in a /v1/systemone request does not change the context.

Measurements and limitations

The selection panels helped choose the files and do not establish performance on unseen inputs. Same-answer rates measure agreement with the original, including its mistakes. Public files can perform better; see the full comparisons on the card. Measurements used llama-server, not Ollama, and Ollama memory use was not measured.

Full measurements and test conditions.

Licence

Apache-2.0 (licence). See the Hugging Face repository for LICENSE and NOTICE.

These files keep improving.